Table Of Content

A model for separatrix splitting near multiple resonances M. Rudnev and V. Ten ∗ † 5 0 January 13, 2004 0 2 n a J Abstract 3 1 We propose a model for local dynamics of a perturbed convex real-analytic Liouville- integrable Hamiltonian system near a resonance of multiplicity 1+m, m 0. Physically, the ] model represents a toroidal pendulum, coupled with a Liouville-integrab≥le system of n non- S linear rotators via a small analytic potential. The global bifurcation problem is set-up for the D n-dimensional isotropic manifold, corresponding to a specific homoclinic orbit of the toroidal . h pendulum. The splitting of this manifold can be described by a scalar function on an n-torus, at whose kth Fourier coefficient satisfies the estimate m O e ρkω kσ , k Zn 0 , [ (cid:16) − | · |−| | (cid:17) ∈ \{ } 1 whereω Rn isaDiophantinerotationvectorofthesystemofrotators;ρ (0,π)andσ >0are 2 v ∈ ∈ theanalyticityparametersbuiltintothemodel. Theestimate,undersuitableassumptionswould 8 generalize to a general multiple resonance normal form of a convex analytic Liouville integrable 0 2 Hamiltonian system, perturbed by O(ε), in which case ωj ∼ √ωε, j =1,...,n. 1 0 1. Introduction and main result 5 0 / The main objective of this paper is to create a template to extend the theory for exponentially h small separatrix splitting in Liouville near-integrable Hamiltonian systems near simple resonances, t a i.e. resonances of multiplicity 1, to the case of multiple resonances, of multiplicity 1+m, m 0. m ≥ The interest in such a theory is dictated by the fact that the normal form theory and Nekhoro- : v shev theorem, resulting in exponentially long time stability (see e.g. [12]) is well developed for i X resonances of all multiplicities, whereas the exponentially small splitting phenomenon, resulting in r similar exponents, has been quantitatively studied so far only in the special case of multiplicity one a resonances. It is well known (see e.g. [2]) that a convex analytic Liouville near-integrable Hamiltonian system, with Hamiltonian H(p,q) = H (p)+εH (p,q), (1.1) 0 1 where (p,q) Rn+1+m Tn+1+m are the action-angle variables on T Tn+1+m, where T = R/2πZ, ∗ ∈ × can be localized in the action space near a multiplicity 1+m resonant action value p . Namely 0 AMS subject classification 70H08, 70H20 ∗Contact address: Department of Mathematics, Universityof Bristol, University Walk, Bristol BS8 1TW, UK; e-mail: [email protected] †Contact address: Department of Mathematics, Universityof Bristol, University Walk, Bristol BS8 1TW, UK; e-mail: [email protected] 1 suppose p is such that the kernel of the scalar product DH (p ),k , k Zn+1+m is some 1+m 0 h 0 0 i ∈ dimensionalsublatticeinZn+1+m. Inthiscase, withoutlossofgenerality onecanrenderDH (p )= 0 0 (ω,0) Rn+1+m. In addition, we further assume that ω Rn is Diophantine, i.e. ∈ ∈ k Zn 0 , k,ω ϑ k τn, (1.2) − ∀ ∈ \{ } |h i| ≥ | | for some ϑ > 0 and τ n 1 (for n = 1 this obviously boils down to ω = 0). n ≥ − 6 After a canonical change of variables, preserving the phase space bundle structure and time scaling the Hamiltonian (1.1) can be cast into the following normal form: ω 1 H (p,q) = ,ι + p,Q p +U(x ,...,x ) + [f (q)+ p,g (p,q) ], (1.3) nf h√ε i 2h nf i 0 m nf h nf i where p = (ι,y ,y ,...,y ) Rn+1+m, q = (ϕ,x ,x ,...x ) Tn+1+m, Q is a constant sym- 0 1 m 0 1 m nf ∈ ∈ metric matrix, and the pair (f ,g )= O(√ε) can betreated as a perturbation when ε (suppressed nf nf inthelatter notations) is small enough. Furtherin thepaper,the boldtypeface marks then+1+m dimensional quantities. If one truncates the normal form Hamiltonian H by dropping the terms f and g in the nf nf nf formula (1.3), the action ι is flow-invariant. For ι = 0, one can separate a natural system of 1+m degrees of freedom, whose Hamiltonian can be written as K(y ,...y )+U(x ,...x ), (1.4) 0 m 0 m whereK(y)isasymmetricpositivedefinitequadraticforminy R1+m andU(x)–ascalarfunction ∈ on T1+m. In the sequel, saying that some function is a “function on a torus” implies 2π-periodicity of this function in corresponding variables. In the simple resonance case m = 0, one can show that inherent in the normal form dynamics is the exponentially small separatrix splitting phenomenon, see [12], [15]. In order to show how the exponentially small splitting theory can be built in the multiple resonance case m 1, let us consider a simple model, which generalizes the so-called Thirring ≥ model for a simple resonance, see [9]. Namely, we study the following model Hamiltonian: n 1 H (p,q)= ω,ι + ι2+H (y ,...,y ,x ,...,x )+µV(ϕ,x ,...,x ), (1.5) µ h i 2 j 1+m 0 m 0 m 0 m Xj=1 where ω is Diophantine, V is a real-analytic function on Tn+1+m and µ is a small parameter. Specifically, for some strictly increasing sequence of positive reals l ,l ,...l , let H (y,x) have 0 1 m 1+m the following form: H (y,x) = K (y)+U (x), 1+m 1+m 1+m K1+m(y0,y1,...,ym) = 21 mi=0 yl2i2, (1.6) i P m m U (x ,x ,...,x ) = l cosx 1 . 1+m 0 1 m i=0 i j=i j − (cid:16) (cid:17) P Q Geometrically, the natural system (1.6) can be visualized as a “toroidal” pendulum, i.e. a particle of unit mass, confined to move on the surface of a ”vertically standing” in R2+m torus of dimension 2 1+m, with principal radii l ,...,l , under the influence of gravity with the free fall acceleration m 0 equal to 1. Mechanically, the case m = 1 can be realized as a double pendulum, whose shorter arm of length l is attached to the terminal point of the longer arm of length l and moves in a circle, 0 1 which rests upon the longer arm. On the energy level H 1 (0), the origin O = (0,0) is a single hyperbolic fixed point, with the 1−+m characteristic exponents i 1 λi = v lj, i = 0,...,m. (1.7) l u iuXj=0 t Suppose the sequence l grows rapidly enough to ensure j λ > max(λ ,...,λ ). (1.8) 0 1 m In addition, we have to assume that m inf (cid:12)λ0 kjλj(cid:12)> 0, (1.9) k Zm(cid:12) − (cid:12) ∈ + (cid:12) Xj=1 (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) where Z denotes non-negative integers. (cid:12) (cid:12) + The origin is connected to itself by a family of homoclinic orbits, in fact there exist homoclines representing each homotopy class on T1+m for the geodesic flow generated by the corresponding Jacobi metric, degenerate at x = 0, see [3]. Some of these homoclinic orbits, or separatrices, are patently obvious: let xj = yj = 0, j 0,...,m i , xi(t) = 4arctane±λit. These orbits ∀ ∈ { } \{ } correspond to homoclinic geodesics forming the basis of the fundamental group of the torus T1+m (modulo the sign in the exponential which bears witness to the reversibility of H , identifying 1+m the upper or lower separatrix branch, where y retains its sign). Consider the orbit with i = 0, call i it γ. This orbit leaves and arrives back to the fixed point in the maximum expansion/contraction direction, correspondingto the Lyapunov exponentλ . Inorder to take both of theorbit’s branches 0 into account, let us represent γ as follows: γ = x = ... = x = y = ... = y = 0, y = 2sin(x /2) ψ(x ), x (0,2π) (2π,4π) . 1 m 1 m 0 0 0 0 { ≡ ∈ ∪ } (1.10) Observe that the existence of the two branches of γ, on each of which y retains its sign, is reflected 0 by 2π-antiperiodicity of the “separatrix function” ψ: ψ(x ) = ψ(x +2π). To reflect this fact, it 0 0 − will be further convenient to deal with x T R/4πZ rather than T = R/2πZ. In particular, 0 2 ∈ ≡ addition of values x is further meant to be mod(4π). Clearly, the orbit γ belongs to both the unstable and the stable 1+m dimensional invariant u,s Lagrangian manifolds W of the fixed point at the origin. If m 1, the flow of the Hamiltonian O ≥ H is non-integrable1. 1+m u,s Global geometry of the manifolds W is complicated. Locally near the origin however, the O u,s u,s germs W of the manifolds W are diffeomorphic to m+1 disks, tangent at O to the unstable O,loc O and stable manifolds of the flow, linearized near the origin. 1The flowof H1+m should possess noglobal analytic first integral other than theenergy,unless K1+m is diagonal and U1+m separated, see [5]. Transversality of the intersection of the manifolds WOu,s along γ (to be shown) is in turn an onset for non-integrability, see [6]. For thegeneral variational approach to homoclinic trajectories in natural systems see [3]. 3 u,s We shall further show that γ arises as a transverse intersection of the manifolds W . Let us O u,s call W the localizations of these manifolds in the neighborhood of γ. As the orbit γ takes off γ from/arrives at the fixed point in the maximum expansion/contraction direction, it itself turns out tobehyperbolicwithinthemanifoldsWu,s. Indeed,onthe“vertical torus”in R2+m, thecoordinate γ directions x ,...,x are the main curvature directions away from γ. 1 m Let us further change the notations (x ,y ) to (x,y), (x ,...,x ) to z and (y ,...,y ) to z¯and 0 0 1 m 1 m restrict z r for some 0 < r < 1. Then 0 0 | | ≤ y2 m z¯2 i z2 H1+m(y,z¯,x,z) = 2l2 +l0(cosx−1)+ 2li2 −l0cosx+ lj 2i +O4(z;x), (1.11) 0 Xi=1 i Xj=1     The semicolon in the symbol O (z;x ) means that the term in question is O( z 4), uniformly in x , 4 0 0 k k standing further for the Euclidean norm, to be used intermittently with the sup-norm . This k·k |·| notational convention will be used further on, the parameters following the semicolon often being omitted. Before formulating the main result, let us give some geometric description of what we are going to claim. Lifted into the phase space of the truncated Hamiltonian H , where µ = 0, the orbit µ γ gives rise to an isotropic n+1 dimensional invariant manifold, which is topologically a cylinder over the n-torus. Let us denote this cylinder as C . Along C , there intersects – degenerately in n O O directions correspondingto therotators’ variable ϕ –a pair of invariant Lagrangian manifoldsWu,s, O both containing an invariant whiskered n-torus T , located at (p,x,z) = (0,0,0). On the torus O itself, the truncated flow is quasiperiodic, with the Diophantine frequency ω. Owingto the fact that C has two branches, plus the fact that the trajectories on C are bi-asymptotic to the invariant O O torus T , we shall technically refer to C as a “bi-infinite bi-cylinder”, yet tending to avoid this O O rhetoric, as much as possible. We study how the presence of the coupling term V in (1.5) is to affect the above described geometric structure and obtain qualitative estimates for the degeneracy removal effect. As far as the Hamiltonian H is concerned, the condition (1.8) results in local hyperbolicity of the orbit 1+m u,s u,s γ within the manifolds W (recall that their localizations near γ are denoted as W ). I.e. the O γ u,s u,s germs W will becontained in the closure of W for the unstable/stable manifolds respectively. O,loc γ Let us denote Wuγ,s ∼= Tn Wγu,s, the lifting of the manifolds Wγu,s into the phase space of the truncated Hamiltonian H ,×when µ = 0. The manifolds Wu,s can berepresented by their generating µ γ functions Su,s(x,z) as graphs over the configurations space variables (ϕ,x,z), where ϕ Tn, z < r γ ∈ | | (for some small enough r to be determined) and x T (2π δ,2π +δ) = [ 2π +δ,2π δ], for 2 ∈ \ − − − some positive δ < 1. Then in the perturbed problem we are going to prove the existence of Lagrangian manifolds Wu,s representing the analogs of the manifolds Wu,s, as far as the Hamiltonian H is concerned. γ µ Moreover, the homoclinic, or bi-infinite cylinder C in the truncated system will give rise to a O pair of semi-infinite cylinders Cu,s in the perturbed system, each perturbed cylinder containing the invariant whiskered torus T, the cylinders Cu,s themselves being contained in the Lagrangian manifolds Wu,s respectively. The phase trajectories on Wu will approach Cu in negative time; in turn the trajectories on Cu, in negative time will approach T at a faster rate. Similar orbit behavior will occur on Ws and Cs in positive time. This is the content of the structural stability theorem, Theorem 2 further in the paper. Moreover, the perturbed Lagrangian manifolds Wu,s can be represented as graphs over the con- figurationspacevariables, by addingtotheunperturbedgenerating functionsSu,s(x,z) respectively, γ 4 some quantities Su,s(ϕ,x,z), which are both O(µ). Then let µ Su,s(ϕ,x,z) = Su,s(x,z)+Su,s(ϕ,x,z) (1.12) γ µ denote the generating functions of the perturbed manifolds Wu,s, respectively. Inordertofindthesegeneratingfunctions,weshalldescribeaseriesofcanonicaltransformations, each of which explicitly takes advantage of the fact that the phase space is a cotangent bundle. In the sequel, any canonical transformation Ψ will be determined by some automorphism a and closed one-form dS on the base space. I.e. all the canonical transformations dealt with herein have the following structure: q = a(q ), Ψ = Ψ(a,S) : (cid:26) p = t(da′)−1p′+dS(q), (1.13) Observe that there is a natural semidirect product structure that on the pairs (a,S), induced by composition. Hyperbolicity of the orbit γ does not suffice to prove Theorem 2 however: we also need a special non-resonance (yet not very restrictive) assumption (1.9) on the stability exponents of H at the 1+m origin, built into the choice of the arm lengths l . The latter assumption appears to be a j j=0,...,m { } very special case of the problem of analytic conjugacy between linearized and non-linear dynamics near a hyperbolic fixed point, see e.g. [13], although for our purposes it suffices separating the dynamics in a single chosen direction only. The whiskered torus T and its local unstable and stable manifold Wu,s are known to survive O O,loc small perturbations without the assumption (1.9), by the theorem of Graff, see [10], [18]. Namely, in the normal form (1.3) as long as ω is Diophantine, U possesses a single non-degenerate absolute maximum, plus the upper left n n minor of the matrix Q is nonzero, there exists a perturbed nf × torus T where the flow is conjugate to the linear flow on its prototype in the truncated system. We further study the splitting of the unperturbed cylinder C . In order to do so, we introduce O the “splitting function” D(ϕ,x,z) = Su(ϕ,x,z) Ss(ϕ,x 2π,z), (1.14) − − which will be well defined for ϕ Tn, x [ 2π +δ, δ] [δ,2π δ] (recall that addition of x is ∈ ∈ − − ∪ − mod(4π)), and z < r. A critical point of D would yield a homoclinic connection to the torus T, | | the gradient dD being the “splitting distance”. u,s As the manifolds W for the Hamiltonian H intersect transversely at z = 0, the critical γ 1+m points of D will lie close to z = 0, and therefore, the magnitude of the splitting of the cylinder C can be evaluated in terms of the derivatives D D(ϕ,x,z) at z = 0, in the properly adjusted O x,ϕ coordinate chart2. The forthcoming Theorem 1 makes these claims precise. In order to formulate the theorem, let us introducesomenotation andsummarizewhatanalyticity propertiesarerequiredoftheperturbation V(ϕ,x,z). For real r,σ > 0 and j = 1,2,... (j = 1 usually being omitted) let Bj d=ef ζ Cj : ζ r , r { ∈ k k≤ } Tj d=ef ζ Cj : ζ Tj, ζ σ . σ { ∈ ℜ ∈ |ℑ |≤ } 2This is the only instant in the argument of this paper, where the built into the model transversality of the intersection of the manifolds Wu,s comes into play. γ 5 For x T , define a conformal map s and some associated quantities as follows: 2 ∈ x dζ s(x)= , χ(s)= ψ[x(s)], e= yχ(s). (1.15) Z ψ(ζ) π Recall, in the model studied ψ(x) = 2sin(x/2). The map s(x) takes (0,4π) to R R + iπ, and ∪ the change (x,y) (s,e) is canonical. The function x(s) is 2πi-periodic and has singularities at → s= πi. ±2 Fix some T 1. By construction of the map s, for any and ρ (0,π/2) any T [T /2,T ], 0 0 0 ≫ ∈ ∈ the quantities x(s),χ(s) are holomorphic functions in the set Πˇ C/2πi, obtained by throwing T,ρ ⊂ out of C horizontal rectangles with half-axes (2T T) (π/2 ρ), centered at πi. Namely, let 0 − × − ±2 Πˇ = Π Π , T,ρ T,ρ T,ρ ∪− Π = s T, s ρ s T, s π ρ s T 2T ; also let (1.16) T,ρ 0 {ℜ ≤ |ℑ |≤ }∪{ℜ ≤ |ℑ − | ≤ }∪{ℜ ≤ − } Πˆ = s C : s T, s ρ . T,ρ { ∈ |ℜ |≤ |ℑ |≤ } The domains Π are further referred to as semi-infinite bi-strips, their size increasing with (T,ρ), T,ρ withρ < π. Bi-stripsΠˇ arebi-infinite,whileΠˆ issimplyanorigin-centeredhorizontalrectangle 2 T,ρ T,ρ in C, with semi-axes (T,ρ). Let C = Tn Π , C = C Bm (1.17) σ,T,ρ σ T,ρ σ,T,ρ,r σ,T,ρ r × × (and in the same fashion Cˇ , Cˇ or Cˆ , Cˆ ) be referred to as complex semi-infinite σ,T,ρ σ,T,ρ,r σ,T,ρ σ,T,ρ,r (bi-infinite or finite) bi-cylinders for C and extended bi-cylinders for the notations C. In qualitative argument, the analyticity indices as well as “bi-” rhetoric are avoided. Let us now quote the main assumption. Assumption 1 Assume the non-resonance conditions (1.2) and (1.9). Suppose the real-analytic function V(ϕ,x,z) is such that V(ϕ,x(s),z) is holomorphic and uniformly bounded by 1 in Cˇ for some initial set (σ ,T ,ρ ,r ) of analyticity parameters. σ0,T0,ρ0,r0 0 0 0 0 The main result of the paper is the following theorem. Theorem 1 Under Assumption 1, take T = T 1 and any positive ρ < ρ , σ < σ , let δ logT. 0 0 0 − ∼ Suppose r <c min[(ρ ρ),(σ σ)], µ < c [rϑ ω 1(σ σ)τn]2, (1.18) 1 0 0 2 − 0 − − | | − for some constants c > 0, determined by the separatrix function ψ as well as the quantities 1,2 n,τ ,m,σ ,T ,ρ ,r ,l ,...l . n 0 0 0 0 0 m i. Some level set of H , with energy O(µ), contains an invariant partially hyperbolic n-torus µ T, where the flow is conjugate to linear, with the rotation vector ω. At the torus T, there intersects a pair of isotropic manifolds Cu and Cs, which are contained respectively inthe global unstable and stable manifolds of T. The manifolds Cu and Cs are contained respectively in a pair of Lagrangian manifolds Wu,s, which are graphs of closed one-forms, with the generating functions Su,s(ϕ,x,z), as in (1.12) respectively, such that the quantities Su,s(ϕ,x(s),z) are µ holomorphic and uniformly bounded by O(µ) for (ϕ,x,z) C . The cohomology classes σ,T,ρ,r ∈ ξu,s H1(Tn,R) = Rn of the one-forms dSu,s are equal to each other. ∈ ∼ 6 ii. The distance between the manifolds Wu,s can be measured by the exact one-form dD, defined by (1.14). There exist a coordinate chart (ϕ,x,z ) Tn [δ,2π δ] Bm, obtained by a ′ ′ ′ r ∈ × − × near-identity change of variables (ϕ,x,z ) = a(ϕ,x,z) from the original coordinates in (1.5), ′ ′ ′ such that in the chart (ϕ,x,z ), the function D satisfies the following PDE: ′ ′ ′ ∂D ∂D ψ(x) + ω, + z ,L[D] = 0, (1.19) ′ ′ ′ ∂x h ∂ϕ i h i ′ ′ where L is a linear first order differentiation operator. ′ iii. In the above chart, namely for (ϕ,x,z ) Cˆ , the function D(ϕ,x,z ) is bounded by ′ ′ ′ σ,T,ρ,r ′ ′ ′ ∈ O(µ). Let D(ϕ,x,z ) = D (x,ϕ)+O(z ). Then the quantity D (x,ϕ) can be written as a ′ ′ ′ 0 ′ ′ ′ 0 ′ ′ 2π-periodic function on Tn of α = ϕ ωs(x), i.e. D (x,ϕ) = S(α). (1.20) ′ ′ 0 ′ ′ − iv. The manifolds Wu and Ws intersect at least 2n+2 orbits, biasymptotic to T. Let us show for the moment that the conclusions (iii) and (iv) of Theorem 1 are straightforward consequences of (i) and (ii). Indeed, (1.19) implies that D (x,ϕ) has to satisfy the linear PDE 0 ′ ′ ∂D ∂D 0 0 ψ(x) + ω, = 0, ′ ∂x h ∂ϕ i ′ ′ and ψ(x) ∂ = ∂ , where s(x) comes from (1.15). Then D (x,ϕ) = S(α) follows, as the form ′ ∂x′ ∂s ′ 0 ′ ′ dD is exact, i.e. D (x,ϕ) is 2π-periodic in ϕ Tn. It follows that the set of the critical points 0 0 ′ ′ ′ ∈ of the function S(α) in the coordinate plane z = 0 determines the trajectories, biasymptotic to ′ the torus T. The minimum number n + 1 of critical points of S is the Ljusternik-Schnirelmann characteristic of Tn, which equals n +1. It gets doubled in the statement (iv), because one can restrict x [ 2π+δ, δ], considering the lower separatrix branch and replicate the statement (ii). ∈ − − Theorem 1 has an immediate corollary, implying the estimate claimed in the Abstract and exponential smallness of the splitting distance if ω ω for a small ε. We do not elaborate → √ε on various parameter relations in this paper (they can be all made to depend on ε in order to approach the lower bounds’ problem for the exponentially small splitting distance) as the situation here would be the same as it is in the simple resonance case, regarding which see e.g. [4] and the references therein. The willing reader may synthesize these relations using the fact that the parameter relations in the forthcoming technical statement Theorem 2 are supposedly optimal. ′ Going carefully through the latter theorem, one can derive what precisely the symbols O(µ) imply regarding the other parameters of Theorem 1. Corollary 1.1 For k Zn 0 , the Fourier coefficients S for of the function S(α) satisfy the k ∈ \{ } estimate S O(µ) e ρkω kσ. (1.21) ′k − | · |−| | | | ≤ · If (ϕ,x,z) are the original coordinates and (ϕ,x,z ) = a(ϕ,x,z) is the change of variables, de- ′ ′ ′ ′ scribed by Theorem 1 (ii), then for all (ϕ,x,z) Tn [δ,2π δ] [ r,r]m, one has a uniform ∈ × − × − bound D a(ϕ,x,z) O(µ) e ρkω kσ. 0 ′ − | · |−| | | ◦ | ≤ · k ZXn 0 ∈ \{ } 7 Observe that Theorem 1 is also valid in the simple resonance case m = 0, when the manifolds C and Cu,s are Lagrangian, rather than isotropic. The simple resonance case has been exposed in detail in [15], see also the references contained therein. In order to keep the ideas clear and not to rival the latter reference length-wise, intermediate steps in deriving many estimates in this paper are omitted, no intermediate analyticity parameters are explicitly introduced, and the constants c 1,2 can get smaller from one statement to another. The reader is often referred to [15], as many aspects of the routine here mimic the m = 0 case. The proof of how Corollary 1.1 follows form Theorem 1 can also be found in the latter reference, as well as [17], [12]. Remark. IntheperturbationV(ϕ,x ,...,x )in(1.5),letV (ϕ,x ) =V(ϕ,x ,0...,0),andsuppose 0 m 0 0 0 it vanishes if x =0. Then one can formally set up the Melnikov integral 0 + M(α) = ∞V (α+ωt,x (t))dt, α Tn, (1.22) 0 0 Z ∈ −∞ and ask whether this quantity is a bona fide C2 approximation for S(α). (A routine calculation shows that setting ω ω results in the Fourier coefficients M , equal to the right-hand side of → √ε k the bound (1.21), with ρ = π.) We do not study this undoubtedly important issue here (see. e.g. 2 [4], [12] for thorough discussion). However, the theory developed further suggests that there are no extra difficulties arising in this respect in the multiple versus the simple resonance case. In particular, the “easy” case n = 1 involving no small divisors should be similar in this respect to the Thirring model, cf. [9]. Remark. Note that contrary to the simple resonance case, where there exists a large body of literature on exponentially small bounds for the splitting of separatrices, see e.g. [12] and the references therein, multiple resonances have been usually approached via the normal forms, alias averaging method. The latter technique (see [12], [14]) is not very explicit geometrically, however as [14] points out, it does enable one to obtain exponentially small upper estimates with sharp constants, which come from dynamical considerations regarding the analyticity domains, if not to relate these estimates directly to the splitting of separatrices. 2. Unperturbed system analysis In this section, for the sake of clarity, we confine ourselves to the case m = 1 only; the extension to m > 1 is transparent. Thus, in this section, let l = 1, l = l > 1. Let us further assume that l is 0 1 such that λ = √1+l < 1 and for k Z l ∈ + λ kλ 1 , k Z , (2.1) + | − | ≥ 10 ∀ ∈ a particular case of the condition (1.9). The Hamiltonian H given by (1.11), for m = 1 turns into 1+m y2 z¯2 z2 H (y,z¯,x,z) = +(cosx 1)+ (l+cosx) +O (z;x). (2.2) 2 2 − 2l2 − 2 4 Let us further change y ψ(x)+y, recall that ψ(x) = 2sin(x/2). Clearly, the choice of the → ± sign as + corresponds to localization near the orbit γ as part of the unstable manifold of the fixed 8 point O, while the sign would imply doing it near γ as part of the stable manifold. Let us call − the resulting Hamiltonians H as follows: 2, ψ ± y2 z¯2 z2 H (y,z¯,x,z) = yψ(x)+ + (l+cosx) +O (z;x). (2.3) 2,±ψ ± 2 2l2 − 2 4 Observe that the Hamiltonians H have resulted from H after the canonical changes with the 2, ψ 2 ± generating functions x S = ψ(ζ)dζ, (2.4) ±ψ ±Z 0 with the + sign for the stable and the sign for the unstable manifolds, respectively. Further − calculations will be quoted for mostly H H only. 2,+ψ 2,ψ ≡ The Hamiltonian H is now a function of x T = R/4πZ, rather than T, with two singular 2,ψ 2 ∈ points, where x= 0,2π. To identify them H , retains a symmetry: 2,ψ H (y,z¯,x,z) =H (y+2ψ(x),z¯,x+2π,z). (2.5) 2,ψ 2,ψ This symmetry was called the “sputnik property” in [15] and was used to validate the analogue of the claim ξu = ξs of Theorem 1 in the case of a simple resonance. Here, as we are dealing with the specific model, we will not use this property explicitly to prove this claim, but rather will later compare the pair H (which is anyway tantamount to the same trick, used in the beginning of 2, ψ ± section 4). u,s To identify the unstable/stable manifolds W of the orbit γ, we will be looking along the γ x-axis at the unstable/stable manifold of the singular point at x = 0 for the Hamiltonians H , 2, ψ ± respectively. This proves convenient and suggests that in general the regular description of the u,s manifolds W is likely to fail in the neighborhood of x =2π. γ Let us linearize the flow of the Hamiltonian H near the orbit γ, whereupon x(t) x (t) = 2,ψ 0 ≡ 4arctanet. For the infinitesimal increments (xˆ,yˆ,zˆ,zˆ¯), one gets the system of equations ± xˆ˙ = Dψ[x (t)]xˆ+yˆ, yˆ˙ = Dψ[x (t)]yˆ, (2.6) 0 0 − zˆ˙ = ˆz¯/l2, ˆz¯˙ = (l 1+tanh2t)zˆ. (2.7) − The tangent space to Wu at the points on γ will be spanned by the vectors – solutions of the latter γ system of linear ODEs, which vanish as t . → −∞ The two pairs of equations (2.6), (2.7) are uncoupled. As far as (2.6) is concerned, there is an obvious solution xˆ(t) = x˙ (t) 1/cosht, yˆ(t) = 0, which vanishes at both t . I.e. one 0 ∼ → ∓∞ tangent direction to Wu at a point (x,z,y,z¯) = (x,0,0,0) is always (1,0,0,0), in the direction γ collinear with γ itself. Equations (2.7) will clearly have no solutions vanishing at both t , as the coefficients therein ±∞ retain their sign for all t. However the system certainly does have a solution (zû(t),zˆ¯u(t)), defined for t T for some T 1, which as t approaches the trivial (zˆ,ˆz¯) (0,0) (as well as 0 0 ≤ ≫ → −∞ ≡ another solution, which is defined for t T and vanishes at t + ). 0 ≥ − → ∞ To construct the unstable solution (zû(t),zˆ¯u(t)), one may set d = ψ(x) d , 1 tanh2t = cosx dt dx − and construct the germ of the solution in question locally as a Taylor series in x near x = 0; by linearity oftheequations(2.7)andboundednessoftheircoefficients, thecontinuation ofthesegerms over a finite time interval does not itself pose any problem3. 3As a matter of fact, equations (2.7) represent a second order linear ODE of generalized Legendre type, whose general solution can befound explicitly in terms of the associated Legendre functions. 9 Observethatgiven(zû(t),zˆ¯u(t)),onecanlet(zˆs(t),zˆ¯s(t)) = (zû( t) 2π, ˆz¯u( t))fortheresult − − − − ofthesimilar procedurewithrespecttotheHamiltonian H . For not [ T ,T ]can thevectors 2, ψ 0 0 (zû(t),zˆ¯u(t)) and (zˆs(t)+2π,zˆ¯s(t))beparallel, or there woul−dexist a solu∈tio−n of (2.7), biasymptotic to zero. For the Hamiltonian H from (2.2), the existence and transversality of the intersection 2 u,s along the orbit γ of a pair of manifolds W (defined in the neighborhood of γ) essentially follow. γ A quantitative statement of this fact is to be given shortly. So far observe by comparing the coefficients in the linear equations (2.7) retain the sign, both vectors (zû(t),zˆ¯u(t)) and (zˆs(t),zˆ¯s(t)) in the (z,z¯) plane never have a slope too close to horizontal. More precisely, their slope in absolute value will be contained in the interval [l√l 1,l√l+1], which can be seen from (2.10) below. − Let us show how the manifold Wu can be constructed, the analysis for Ws gets modified in the γ γ obvious way. Make a change z¯ z¯+λ (x)z, where λ (x) determines the direction of the solution u u → vector, vanishing at x = 0 (i.e. t ). The quantity λ (x) [l√l 1,l√l+1] is well defined u → −∞ ∈ − for x T (2π δ,2π +δ) for some 0 < δ < 1, where δ lnT . 2 0 ∈ \ − ≈ The change z¯ z¯+ λ (x)z is not canonical, to make up for it one also has to change y u → → y+ 1dλu(x)z2. In other words, this is a canonical change with the generating function 2 dx 1 Su (x,z) = λ (x)z2. (2.8) γ,0 2 u Then the Hamiltonian H in (2.3) transforms to 2,ψ y2+l 2z¯2 1 H (x,y) = yψ(x)+ − + λ˜ (x)z2 +l 2λ (x)zz¯+yO (z;x)+O (z;x), (2.9) 2,u u − u 2 4 2 2 where dλ (x) λ˜(x) = ψ(x) u (l+cosx)+l 2λ2(x). (2.10) dx − − u It follows by construction – or directly from (2.7) – that λ˜ (x) 0. Thus the quantity Λ (x) u u ≡ ≡ l 2λ (x) multiplying zz¯ in (2.9) is always positive, never exceeding λ = Λ (0); recall that λ < 1, − u u also cf. (2.1). Finally, the last two terms in (2.9) can be regarded as a perturbation, provided that r is small enough. The phase space of the Hamiltonian H is T ((T (2π δ,2π +δ)) [ r,r]). If there were 2,u ∗ 2 \ − × − no two last terms in (2.9), the manifold Wu would be given by the zero section (y,z) = (0,0) of γ the bundle. However, for small r, the last two terms in (2.9) can be regarded as a perturbation and dispensed with, owing to the following lemma. Lemma 2.1 Given ρ < ρ , T T 1, there exists a constant c > 0, depending only on the 0 0 1 ≤ − parameter set (ρ ,T ,r ) and λ, such that for r < c (ρ ρ), there exists some reals δ,κ = O(1) in 0 0 0 1 0 − (0,1) and a canonical near-identity transformation Ψu, such that the Hamiltonian (2.9) can be cast r into the following normal form: H (y,z¯,x,z) = yψ(x)+Λ (x)zz¯+O (y,z¯), (2.11) γ,u u 2 valid for y , z¯ κ, z r and x such that x [ 2π+δ,2π δ] and s(x) Π . T,ρ | | | | ≤ | | ≤ ℜ ∈ − − ∈ The transformation Ψu, for p = (y,z¯) and q =(x,z) can be written in the following form: r q = q + bu(q ), Ψur = Ψur(bur,Sur) : (cid:26) p = t[id+dbur′(q′)]−1p′ + dSrur(q′), (2.12) where the quantities bu(x,z) and Su(x,z) are both O (x + z ). r r 2 | | | | 10